SMOTE and ABC optimised RBF network for coping with imbalanced class in EEG signal classification

Abstract

This paper proposes a novel approach for coping with imbalanced class problem by combining the best attribute of synthetic minority over-sampling technique (SMOTE) and artificial bee colony optimised radial basis function neural networks to identify epileptic seizure from electroencephalography (EEG) signal. EEG is the recording of electrical activity in brain. Careful analysis of these recordings can provide valuable information and understanding the mechanisms of several brain disorder diseases such as epilepsy. Since epileptic seizures occur irregularly and unpredictably, automatic seizure detection in EEG recordings is highly required. We have used discrete wavelet transform (DWT) technique for extraction of potential features from the signal. For classification of these signals into two classes, we have trained the RBFN by a modified version of ABC algorithm (MABC). In this work, we realise, this two class classification problem is highly imbalanced i.e., the instances in one class known as majority class outnumber the instances of other class called the minority class. The SMOTE is first applied to generate synthetic instances in the positive class to balance the training data set. Using the resulting balanced dataset, the MABC optimised RBF network is then constructed to identify the epileptic seizure.